Artificial intelligence-based express industry personnel safety management method and related device

Abstract

The application relates to an express industry personnel safety management method based on artificial intelligence and related devices, which comprises the following steps: generating training content according to the post-related data of employees, generating a training effect evaluation result according to the completion of the training content by the employees, and granting the employees on-duty permissions in response to the evaluation result reaching a preset evaluation threshold; in response to the employees having the on-duty permissions and entering an on-duty state, performing rule violation identification on the work video of the employees to obtain a rule violation judgment result, and updating the rule violation points of the employees according to the rule violation judgment result; in response to the rule violation points meeting a reinforcement training triggering condition, generating a reinforcement training instruction and returning it to the training link, driving the regeneration and re-execution of the training content for the employees, and locking the on-duty permissions of the employees during the period. The application can construct a feedback link from the supervision link to the training link, and support continuous rectification of similar rule violation behaviors.

Description

Technical Field

[0001] This application relates to the field of management systems, and in particular to an artificial intelligence-based method and related apparatus for personnel safety management in the express delivery industry. Background Technology

[0002] With the rapid development of the e-commerce industry, the volume of express delivery business has continued to climb, and the number of employees has been expanding. Consequently, express delivery companies have a growing need for safety management of their frontline employees. Frontline employees in express delivery companies typically work in various positions, including sorting, delivery, and security checks, and their work environments encompass multiple locations such as branches, sorting centers, and delivery routes. Safety risks are dispersed and diverse, therefore, express delivery companies need to implement various safety management measures in their daily operations, including onboarding training, on-the-job violation supervision, and emergency response procedures.

[0003] Currently, personnel safety management in the express delivery industry largely adopts a segmented management model, which involves deploying independent management systems or manual control processes for different management stages such as training, supervision, and emergency response. This segmented model has a core problem: data is not shared between management stages, and processes are not interconnected. This means that employee safety performance data collected in the back-end stages cannot be used to inform management strategies in the front-end stages. Specifically, the training content received by employees upon joining the company is usually generated based on a standardized job template, and once distributed, it is decoupled from the employee's actual performance after starting work. After employees are on duty, existing methods of monitoring violations mainly rely on manual inspections or simple video recordings. Even if violations are identified, the results are only recorded and cannot be fed back to the training stage to drive retraining for that employee. Similarly, operational data such as violations, emergency events, and weather risk responses that occur during employee work hours stop at their respective independent management systems and do not affect subsequent job qualification determinations.

[0004] This leads to the problem that once employees pass onboarding training and obtain their job qualifications, even if they repeatedly commit similar violations during their employment, the existing management model cannot automatically trigger targeted retraining and reassessment. Employees' safety awareness and operational standards are difficult to continuously strengthen, and similar violations are prone to recurrence. If managers wish to retrain employees who repeatedly violate regulations, they often need to rely on manually initiating assessment processes, manually locking job access, and manually adjusting training content. This process is cumbersome and slow to respond, making it difficult to support the refined safety management needs of large-scale operations in the express delivery industry. Furthermore, the lack of data sharing between different stages makes it difficult for companies to continuously optimize training programs and supervision strategies based on employee performance, resulting in insufficient overall and adaptive safety management capabilities. Summary of the Invention

[0005] To address the problems of the current segmented management model for personnel safety management in the express delivery industry, which results in data not being shared between management links, the inability of backend supervision results to feed back into frontend training, leading to repeated violations by employees that are difficult to eradicate, and insufficient overall management and adaptability, this application provides an artificial intelligence-based method and related device for personnel safety management in the express delivery industry.

[0006] Firstly, this application provides an artificial intelligence-based method for personnel safety management in the express delivery industry, which adopts the following technical solution: An AI-based method for personnel safety management in the express delivery industry includes the following steps: S1. Generate training content for employees based on their job-related data, and generate training effectiveness evaluation results based on employees' completion of the training content. When the training effectiveness evaluation results reach the preset evaluation threshold, grant employees the right to take up their posts. S2. In response to an employee having the authority to work and entering the on-duty work state, perform violation identification on the employee's work video, obtain the violation judgment result, and update the employee's violation points based on the violation judgment result; S3. In response to the violation points meeting the conditions for triggering enhanced training, an enhanced training instruction is generated and sent back to S1. The enhanced training instruction is used to drive the regeneration and re-execution of training content for employees, and locks the employee's work permissions during the regeneration and re-execution. This allows the violation judgment result to be applied to the regeneration of training content through the enhanced training instruction, thus constructing a feedback link from S2 to S1.

[0007] By adopting the above technical solution, the violation judgment results generated in the S2 violation supervision link are accumulated through violation points. When the enhanced training threshold is triggered, the enhanced training instruction is sent back to the S1 training link, so that S1 regenerates training content for the same employee and performs the assessment again. This establishes an automated feedback link between the violation supervision link and the training link. At the same time, by locking the employee's on-duty access, the employee cannot generate new violation data during the feedback period. The employee's violation performance and training content remain dynamically coupled during the operation of the method. Employees who repeatedly violate the rules can be continuously corrected. The overall safety management and adaptive capability gradually increase with the continuous accumulation of employee on-duty data.

[0008] Optionally, S1 includes sub-steps S11-S14: S11. Perform structured parsing of employees' identification information and professional qualification information to obtain structured employee files; S12. Match the job assignment information and corresponding work processes and safety risk points from the job characteristic database, and output a job training needs list; S13. Input the job training needs list and historical accident case data into the course generation model, output the training content, and generate a matching exam based on the training content and historical accident case data. S14. Collect data on employees' learning time, course completion progress, and learning interaction related to the training content. Combine this data with the results of the accompanying exams and generate training effectiveness evaluation results using a fuzzy comprehensive evaluation algorithm.

[0009] By adopting the above technical solution, the four sub-stages of structured files, job requirement matching, synchronous generation of courses and exams, and comprehensive evaluation of multi-dimensional data are linked together. This ensures that the training content is precisely aligned with the work processes and historical accident experience of the employees' positions. The training effectiveness evaluation results are based on multi-dimensional data generation rather than a single exam score, providing a quantifiable basis for determining the authorization of S1 employees to take up their posts.

[0010] Optionally, S2 includes sub-steps S21-S24: S21. Collect operation videos through visual acquisition devices, including fixed visual acquisition devices deployed in the operation area and vehicle-mounted visual acquisition devices deployed in delivery vehicles; S22. Perform preprocessing on the work video to obtain the preprocessed work video; S23. Input the preprocessed operation video into the violation identification model. The violation identification model is trained using a violation sample set of the express delivery industry. The violation identification model performs feature extraction and classification identification on the preprocessed operation video and outputs information on suspected violations. S24. Verify suspected violation information using preset violation judgment rules, and output violation judgment results, including violation type and violation level.

[0011] By adopting the above technical solution, the four sub-stages of data collection, preprocessing, model reasoning, and rule verification are linked together, enabling fixed vision acquisition equipment and vehicle-mounted vision acquisition equipment to cover two types of operation scenarios: the sorting area of ​​the network point and the delivery link. The suspected violation information output by the model is then verified a second time by the preset violation judgment rules, which reduces misidentification and adds type and level information to the violation judgment results output by S2, supporting the differentiation of the violation score update range of S3.

[0012] Optionally, the violation detection model is built on an object detection network. A channel attention structure is embedded in the feature extraction layer of the object detection network. The channel attention structure is used to recalibrate the weights of the feature channels to obtain enhanced features for violations of small targets.

[0013] By adopting the above technical solution, the channel attention structure assigns differentiated weights to the feature channels output by the feature extraction layer, thereby improving the accuracy of the violation identification model in identifying small-target violations such as not wearing gloves or improperly placing packages.

[0014] Optionally, after S2 obtains the violation determination result, the following tiered response operations are also included: If the severity of the violation indicated by the violation determination result does not reach the preset linkage threshold, a reminder is issued to the employee via the employee's handheld terminal and on-site voice broadcast; If the severity of the violation indicated by the violation determination result reaches the preset linkage threshold, in addition to issuing a reminder to the employee, an intervention reminder is simultaneously issued to the management personnel's terminal, and an equipment linkage command is sent to the work equipment control system to trigger the work equipment control system to brake the corresponding work equipment; The violation determination result, the corresponding work video clip, and the identification information indicating the employee's identity are encapsulated into a violation record entry, and the violation record entry is written to the violation record storage area.

[0015] By adopting the above technical solution, the violation response is divided into two levels, mild and severe, with a preset linkage threshold as the dividing point. Mild violations only remind the employee, while severe violations involve management intervention and braking of the operating equipment. The violation record item retains the violation judgment result and corresponding video evidence in its entirety, providing a basis for subsequent assessment and accountability as well as iterative optimization of the violation identification model.

[0016] Optionally, the method further includes the following steps: S4. Obtain emergency query text from employees, perform semantic parsing on the emergency query text through a large safety model, and combine the emergency database to output the corresponding handling plan suggestions. The emergency database includes a historical emergency event experience database and an emergency plan database.

[0017] By adopting the above technical solutions, employees can obtain targeted handling suggestions through natural language queries when encountering emergencies. The safety big data model generates suggestions based on the historical emergency experience database and emergency plan database, extending the management dimensions covered by the method from daily training and supervision to emergency response to emergencies.

[0018] Optionally, the method further includes the following steps: S5. Obtain meteorological risk data of the work area, determine the warning level based on the meteorological risk data, and generate warning push information for employees based on the warning level.

[0019] By adopting the above technical solutions, meteorological risk information is integrated into the process flow, enabling employees to obtain early warning information before encountering extreme weather conditions. The early warning information complements the on-duty supervision in S2 and the emergency response in S4, covering the risk avoidance dimension before operations.

[0020] Optionally, structured employee profiles are encrypted before being written to the data storage layer.

[0021] By adopting the above technical solutions, employees' sensitive identity information is encrypted and protected during storage, meeting industry data security standards.

[0022] Optionally, the training content includes basic courses, job-specific courses, and case study courses, which are divided into multiple independent learning modules. Each independent learning module supports saving learning progress.

[0023] By adopting the above technical solution, the training content is classified by course type and further broken down into independent learning modules, which is suitable for the scenario where front-line employees are busy and have difficulty concentrating on learning. Employees can complete the training in their spare time.

[0024] Optional evaluation indicators for training effectiveness include training completion rate, exam pass rate, practical compliance rate, and accident rate.

[0025] By adopting the above technical solutions, the evaluation indicators are expanded to include multiple dimensions, including on-the-job performance, so that the training effectiveness evaluation results not only reflect the degree of knowledge mastery during the learning phase, but also reflect the employees' standard execution in actual operations.

[0026] Optionally, the express delivery industry violation sample set includes employee operation violation samples and delivery traffic violation samples. The express delivery industry violation sample set is subjected to data augmentation processing to expand the sample size before being input into the violation identification model for training.

[0027] By adopting the above technical solutions, the sample set covers two typical violation scenarios: operation and traffic. Data augmentation processing expands the sample size in the pre-training stage, improving the violation identification model's ability to generalize to unseen samples.

[0028] Optionally, the violation levels are divided into general violations, serious violations, and major violations, with different violation levels corresponding to different update ranges of violation points.

[0029] By adopting the above technical solution, the violation judgment result is quantified into a differentiated update range of violation points, so that major violations can be quickly promoted to the trigger condition of enhanced training after a single occurrence, while minor violations need to be accumulated multiple times before triggering.

[0030] Optionally, the weights of the violation detection model are quantized in INT8 format, which is suitable for deployment in embedded devices and vehicle terminals.

[0031] By adopting the above technical solutions, the violation detection model can be deployed to edge devices, reducing reliance on cloud computing power and enabling S2 to continue performing violation detection even in network outage scenarios.

[0032] Optionally, violation records can be stored using blockchain technology.

[0033] By adopting the above technical solutions, the process of storing violation records is tamper-proof, providing a credible chain of evidence for employee appeals and corporate compliance tracing.

[0034] Optionally, the security big model is built on the Transformer architecture and pre-trained and fine-tuned using an emergency database.

[0035] By adopting the above technical solutions, the safety big data model is equipped with specialized semantic understanding capabilities for emergency scenarios in the express delivery industry, and the output disposal plan suggestions are aligned with the risk characteristics of express delivery operations.

[0036] Optionally, in response to the fact that the parsing result of the emergency query text by the security big data model does not match a valid item in the emergency database, a prompt for manual emergency command access is generated, and the applicable scenarios and capability boundaries are marked in the output disposal plan suggestions.

[0037] By adopting the above technical solutions, a human safety net is introduced outside the capability boundaries of the large safety model to prevent employees from making incorrect decisions based on uncertain outputs. At the same time, boundary labeling is used to establish cognitive alignment between employees and the model.

[0038] Optionally, the completed emergency response records can be backfilled into the historical emergency event experience database for incremental training of the safety model.

[0039] By adopting the above technical solutions, new experience data generated during the emergency response process is fed back into the historical emergency event experience database, and the safety big model is continuously optimized through incremental training. The emergency response capability of the method evolves as the runtime increases.

[0040] Optionally, the determination of the warning level may also take into account the geographical environment data of the work area.

[0041] By adopting the above technical solutions, the warning level determination is not only based on the meteorological risk data itself, but also combined with the geographical environmental characteristics of the operation area, such as terrain, road conditions, and network distribution, so that the warning levels of different operation areas under the same meteorological risk level can be differentiated.

[0042] Optionally, the warning levels are divided into red, yellow and blue warnings, with different outdoor work restriction policies corresponding to different warning levels.

[0043] By adopting the above technical solution, abstract meteorological risks are mapped into three levels of differentiated outdoor operation restriction strategies, and employees and managers can directly implement corresponding operation adjustment actions according to the warning level.

[0044] Optionally, monitoring strategy adjustment parameters are generated based on the warning level, and these parameters are injected into the violation identification of S2. This allows the identification focus and violation judgment rules of S2 to be adjusted synchronously with the warning level, enabling S5 and S2 to form a cross-linkage.

[0045] By adopting the above technical solutions, the warning level of S5 drives the dynamic adjustment of the violation identification focus and judgment rules of S2, making the violation identification of outdoor delivery under extreme weather conditions more stringent. An automated linkage is formed between the weather warning link and the on-duty supervision link, supplementing the supervision to training feedback link constructed by S3 as another cross-link linkage link.

[0046] Optionally, the enhanced training trigger conditions include at least one of the following: the cumulative value of violation points within a preset statistical period reaches a preset point threshold; the violation level corresponding to the violation points reaches a preset level; or the violation type corresponding to the violation points is a preset high-risk type.

[0047] By adopting the above technical solutions, the training trigger conditions are enhanced to support three complementary judgment dimensions, so that repeated minor violations, single major violations, and specific high-risk violations by the same employee can all trigger enhanced training, and the triggering mechanism covers different violation patterns.

[0048] Optionally, this application also provides an AI-based personnel safety management system for the express delivery industry. This system includes a personnel intelligent training subsystem, a personnel violation supervision subsystem, and an interface adaptation layer connecting the personnel intelligent training subsystem and the personnel violation supervision subsystem. The personnel intelligent training subsystem is configured to: generate training content for employees based on their job-related data; generate training effectiveness evaluation results based on employees' completion of the training content; and grant employees work access permissions in response to the training effectiveness evaluation results reaching a preset evaluation threshold. The personnel violation supervision subsystem is configured to: acquire work videos of employees with work access who are in an on-duty work state; perform violation identification on the work videos; obtain violation judgment results; update employees' violation points based on the violation judgment results; and generate reinforcement training instructions in response to the violation points meeting the reinforcement training trigger conditions. The interface adaptation layer is configured to: transmit the reinforcement training instructions from the personnel violation supervision subsystem to the personnel intelligent training subsystem, driving the regeneration and re-execution of the training content for employees, and locking employees' work access permissions during the regeneration and re-execution process.

[0049] By adopting the above technical solution, and taking the collaboration of the personnel intelligent training subsystem, the personnel violation supervision subsystem, and the interface adaptation layer as the carrier, the method steps of S1 to S3 are mapped to the corresponding subsystems and interface components. The interface adaptation layer serves as the physical carrier of the feedback link, enabling the reinforcement training instructions to be uniformly scheduled and controlled during the transmission across subsystems.

[0050] Optionally, the system also includes a personnel safety emergency subsystem, which is configured to acquire meteorological risk data of the work area, determine the warning level based on the meteorological risk data, and generate warning push information for employees based on the warning level.

[0051] By adopting the above technical solutions, the system expands its meteorological risk response capabilities, with early warning pushes carried out by an independent subsystem, operating in parallel with the training and supervision subsystems.

[0052] Optionally, the system also includes a data storage layer and an algorithm support layer. The data storage layer is configured to store job-related data, training effectiveness evaluation results, violation judgment results, and violation points. The algorithm support layer is configured to support the operation and updating of the violation identification model.

[0053] By adopting the above technical solution, the underlying storage and algorithm resources are decoupled from the business subsystem, enabling the business subsystem to share the same set of data and algorithm assets, and the updates and iterations of the violation identification model do not affect the upper-level business processes.

[0054] Optionally, the data storage layer adopts a distributed storage architecture of relational databases and non-relational databases. The relational database stores structured data such as employee information, training data, violation records and points data, while the non-relational database stores unstructured data such as video data, image data, emergency event data and emergency plan documents. The data storage layer supports data sharding and disaster recovery backup.

[0055] By adopting the above technical solutions, structured data and unstructured data are respectively adapted to the transaction consistency advantages of relational databases and the large-capacity storage advantages of non-relational databases. The sharding and disaster recovery mechanisms ensure the availability of the data storage layer under large-scale employee data.

[0056] Optionally, the algorithm support layer is built on deep learning frameworks, computer vision algorithm libraries, and natural language processing algorithm libraries, and supports online updates and iterations of algorithm models.

[0057] By adopting the above technical solutions, the algorithm support layer uses a general framework and algorithm library as its foundation, enabling the violation identification model, course generation model, and security big model to share a single infrastructure. The online update capability allows algorithm iteration to be performed without downtime.

[0058] Optionally, the interface adaptation layer includes device interfaces, third-party interfaces, and internal interfaces. Device interfaces are used to interface with hardware devices, third-party interfaces are used to interface with external system interfaces, and internal interfaces are used to enable data communication between subsystems. The interface adaptation layer adopts the RESTful protocol and performs encryption processing on data transmission.

[0059] By adopting the above technical solution, the interface adaptation layer is divided into three types of interfaces according to the type of interface object. The standardized protocol and encrypted transmission take into account both compatibility and security, enabling the system to have the expansion capability to interface with the enterprise's existing management system and new hardware devices.

[0060] Optionally, the system deployment adopts a cloud-edge collaborative deployment mode. The cloud deploys the backend management, security big data model and data storage layer, while the edge deploys the violation identification algorithm, visual acquisition device and device linkage. The edge supports offline operation and will synchronize offline data to the cloud after the network is restored.

[0061] By adopting the above technical solutions, computing power and data storage are centralized in the cloud to support large-scale operations, while violation identification and device linkage, which have high real-time requirements, are decentralized to the edge. Offline capabilities enable the system to adapt to scenarios with wide distribution of express delivery outlets and variable network environments.

[0062] Optionally, the system also includes a back-end management subsystem, which is configured to provide functions such as employee management, training management, violation management, emergency management, system configuration, and data statistical analysis, and supports the generation of multi-dimensional safety management statistical reports by time, position, and region.

[0063] By adopting the above technical solution, the back-end management subsystem serves as a unified control entry point, centralizing the data and configurations of each subsystem onto a single operating interface, and generating multi-dimensional reports to provide data views for enterprise security decisions.

[0064] Secondly, the computer device provided in this application adopts the following technical solution: A computer device comprising: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to: Implement the aforementioned AI-based personnel safety management methods in the express delivery industry.

[0065] Thirdly, this application provides a computer-readable storage medium that adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above regarding the AI-based personnel safety management method for the express delivery industry.

[0066] The storage medium stores at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the following: Such as the above-mentioned AI-based personnel safety management method in the express delivery industry.

[0067] In summary, this application includes at least one of the following beneficial technical effects: 1. The violation judgment results generated through the S2 violation supervision process are driven by the accumulation of violation points to send the S3 reinforcement training instructions back to the S1 training process. An automated feedback link is established between violation supervision and training, so that employees who repeatedly violate the rules can be continuously corrected. This eliminates the problem of decoupling supervision results and training content in the traditional segmented management model.

[0068] 2. By introducing detailed methods such as job characteristic matching and fuzzy comprehensive evaluation in S1, and introducing detailed methods such as channel attention structure, graded response, and violation record retention in S2, the training content is precisely aligned with job risks, violation identification is more sensitive to small targets, and violation handling is graded according to severity, thus further ensuring the input and output quality at both ends of the feedback chain.

[0069] 3. By dynamically adjusting the S2 supervision strategy through S4 emergency query, S5 weather warning, and S5 warning level, the management dimensions covered by the method are expanded from the training-supervision two links to emergency response and risk warning. A supplementary cross-link linkage is established between weather warning and on-the-job supervision, which improves the integrity and adaptability of the safety management system in multiple scenarios. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the overall architecture of an artificial intelligence-based personnel safety management system for the express delivery industry, as described in one embodiment of this application.

[0071] Figure 2 This is a schematic diagram of the architecture of the personnel intelligent training subsystem in one embodiment of this application.

[0072] Figure 3 This is a schematic diagram of the architecture of the personnel violation monitoring subsystem in one embodiment of this application.

[0073] Figure 4 This is a schematic diagram of the architecture of the personnel safety emergency subsystem in one embodiment of this application.

[0074] Figure 5 This is a schematic diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0075] The present application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the scope of the application.

[0076] This application provides an AI-based method for personnel safety management in the express delivery industry. The method revolves around employee training and release, on-the-job violation supervision, and the feedback loop between the two. It generates targeted training content based on employee job-related data and grants work access to qualified employees. During an employee's on-the-job period, violation identification is performed on their work videos to obtain violation judgment results, and the employee's violation points are updated based on these results. When the violation points meet the trigger conditions for enhanced training, an enhanced training instruction is generated and sent back to the training stage, driving the regeneration and re-execution of training content for the same employee, while simultaneously locking the employee's work access, thus forming an automated feedback loop between the violation supervision and training stages. The following provides a detailed description of each step and its implementation.

[0077] For ease of description, the core terms and benchmark application scenarios involved in this method are first defined.

[0078] Job-related data refers to a structured collection of data related to an employee's job identity, job assignment, and job operation standards. Its scope covers employee identification information, professional qualification information, job assignment information, job operation processes, and job safety risk points.

[0079] Operational video refers to a continuous video stream captured in real time by visual acquisition equipment during the employee's on-duty operation. It includes both the operation scenes of employees in fixed operation areas such as sorting stations, security checkpoints, and branch entrances and exits, as well as the driving scenes of delivery personnel in delivery vehicles and road environment scenes.

[0080] The violation determination result refers to the output of the judgment based on the work video to identify employees who violated work specifications or traffic rules, and includes at least two fields: violation type and violation level.

[0081] Violation points refer to the cumulative quantitative record of employee violations. They are maintained independently for each employee in the form of a points account, and the account value is updated according to a preset range when a violation occurs.

[0082] The enhanced training trigger condition refers to the set of conditions that trigger the feedback instruction when the violation points meet the preset criteria. The action after triggering is to drive S1 to re-execute the training release process.

[0083] On-duty access refers to the permission mark that determines whether an employee can enter the on-duty work status.

[0084] The feedback chain refers to the signaling path that ultimately affects the training process after the violation judgment results generated in the violation supervision process pass through intermediate links such as violation point accumulation, determination of enhanced training trigger conditions, and generation of enhanced training instructions.

[0085] As the baseline application scenario throughout this paper, we assume that employee 001 is a deliveryman with employee number E001, working at the Chengdong branch, covering the Chengdong area, and using an electric delivery vehicle. Employee 001 is introduced to this method on their first day of employment and goes through several stages, including training and release, onboarding, violation triggering, feedback and reinforcement training, emergency query, and weather warning response. The S1 training and release stage will be introduced first below.

[0086] S1. Generate training content for employees based on their job-related data, and generate training effectiveness evaluation results based on employees' completion of the training content. When the training effectiveness evaluation results reach the preset evaluation threshold, grant employees the right to take up their posts.

[0087] S1 serves as the entry point for this method, responsible for generating training content based on the individualized characteristics of each employee's position and performing qualification checks. As an example, position-related data can consist of three parts: the employee's identification information, professional qualification information, and job assignment information. Identification information confirms the employee's legal status; professional qualification information determines whether the employee possesses the prerequisite qualifications for a specific position; and job assignment information indicates the specific job type the employee is assigned to, such as sorter, delivery person, security inspector, or branch manager. Taking employee 001 as an example, employee 001's identification information includes name, ID number, and date of birth; professional qualification information includes a driver's license; and job assignment information is delivery person, with a delivery area in the eastern part of the city.

[0088] The input-output closed loop of S1 is as follows: input is job-related data, output is work authorization. This involves two data transformations: training content is generated from the job-related data, and a training effectiveness evaluation result is generated based on the employee's completion of the training content. The training effectiveness evaluation result is a quantifiable comprehensive score, which is compared with a preset evaluation threshold to determine whether work authorization is granted. The preset evaluation threshold is determined by the company's safety management regulations and can be a fixed percentage score (e.g., 80 points) or a multi-dimensional weighted comprehensive score (e.g., 0.85). This embodiment does not limit the specific value of the preset evaluation threshold. When employee 001's training effectiveness evaluation result does not reach the preset evaluation threshold, employee 001 will not be granted work authorization and will need to participate in training again and undergo another evaluation until the evaluation result reaches the preset evaluation threshold.

[0089] Specifically, refer to Figure 2 S1 includes sub-steps S11-S14.

[0090] S11. Perform structured parsing of employees' identification information and professional qualification information to obtain structured employee files.

[0091] S11 employs Optical Character Recognition (OCR) technology to perform structured parsing of employee identification documents. On their first day of employment, employee 001 places their original ID card and driver's license on a scanning device. The device captures the images and sends them to the OCR engine, which identifies the text fields on the documents and maps the results to structured fields according to preset field templates. For example, the ID card recognition result template includes name, gender, ethnicity, date of birth, ID number, issuing authority, and validity period; the driver's license recognition result template includes name, driver's license number, permitted vehicle type, initial license issuance date, and validity period. After recognition and mapping, employee 001's identity fields and professional qualification fields are merged and encapsulated to form employee 001's structured employee file. The structured employee file is organized in key-value pairs, facilitating subsequent sub-steps to retrieve and access information by field name.

[0092] In some embodiments, structured employee profiles are encrypted before being written to the data storage layer. Encryption can employ symmetric encryption algorithms to perform field-level encryption on sensitive fields such as ID card numbers and driver's license numbers, with the key centrally managed by an enterprise key management service. After encryption, the structured employee profiles in the data storage layer are in an encrypted state during the static storage phase. No module without the corresponding key can directly read the plaintext content of sensitive fields, thus protecting the employee's identity and privacy information.

[0093] S12. Match the job assignment information and corresponding work processes and safety risk points from the job characteristic database, and output a job training needs list.

[0094] The job characteristic database is a pre-built database that stores standardized characteristic descriptions of various jobs in the express delivery industry. It includes at least fields such as job title, work process, safety risk points, recommended training focus, recommended training duration, and recommended training frequency. S12 uses employee 001's job assignment information as the search key to match the corresponding entry for the delivery person job in the job characteristic database. It then reads the work process and safety risk point fields under that entry and combines them with other fields to synthesize a job training needs list for employee 001.

[0095] Taking employee 001 as an example, the delivery personnel entries matched from the job characteristic database include the following work process fields: "pickup → loading → delivery → delivery completion confirmation → return trip"; and the safety risk field includes "traffic violations, heatstroke during outdoor high-temperature operations, slippery conditions during heavy rain, delivery conflicts, and vehicle malfunctions". S12 generates a job training needs list for employee 001 based on these fields. The list includes the following: the training focus is on traffic rules, outdoor protection, and emergency response; the recommended training duration is 8 hours; and the recommended training frequency is one refresher training session per quarter.

[0096] S13. Input the job training needs list and historical accident case data into the course generation model, output the training content, and generate a matching exam based on the training content and historical accident case data.

[0097] Historical accident case data is a long-term maintained dataset derived from internal accident records of express delivery companies, publicly available data from industry regulatory departments, and statistical data from third-party safety agencies. It stores fields such as accident type, cause, involved positions, handling process, and corrective measures for past safety incidents in the express delivery industry. S13 takes employee 001's job training needs list and the selected historical accident case data as input and feeds them into the course generation model, which automatically generates training content specifically for employee 001.

[0098] The course generation model is built on the Transformer architecture and trained with a large number of sample pairs of "job training needs list + historical accident case data → high-quality training courses". It has the ability to organize course structure, arrange course content, and embed accident cases based on input data. The training content output by the course generation model can include multiple interconnected course items. Each course item includes fields such as course title, course duration, course text, and supporting material references. Taking employee 001 as an example, the training content output by the course generation model includes multiple course items such as "Basic Traffic Safety for Delivery Personnel", "Electric Vehicle Riding Regulations", "Self-Rescue and Mutual Rescue for Sudden Illnesses During Delivery", and "Coping with Heavy Rain During Delivery", with each item having a course duration between 10 and 20 minutes.

[0099] While delivering training content, S13 also generates a supplementary exam based on the training content and historical accident case data. The exam includes multiple-choice questions, true / false questions, and case analysis questions. The multiple-choice and true / false questions are constructed from knowledge points extracted from the training content, while the case analysis questions select typical accidents from historical accident case data as case backgrounds, requiring employees to answer questions on accident prevention, on-site handling, and post-accident rectification.

[0100] In some embodiments, the training content includes basic courses, job-specific courses, and case study courses, divided into multiple independent learning modules. Each independent learning module supports saving learning progress. The basic courses cover the general principles of safety management and common safety standards in the express delivery industry. The job-specific courses are specifically designed to address the safety risks associated with the employee's position. The case study courses review typical accident cases relevant to the employee's job. Each independent learning module is 10 to 20 minutes long, allowing employees to complete individual modules during breaks in deliveries or while waiting at branch locations. Learning progress for each module is automatically saved by the system, allowing employees to resume learning from their previous point.

[0101] S14. Collect data on employees' learning time, course completion progress, and learning interaction related to the training content. Combine this data with the results of the accompanying exams and generate training effectiveness evaluation results using a fuzzy comprehensive evaluation algorithm.

[0102] S14 collects multidimensional behavioral data on employee 001 during the training process, including cumulative learning time, percentage completion of each course item, accuracy rate of answers to interactive questions within the course, and completion rate of course videos. In addition, S14 also reads employee 001's total score on the accompanying exam and the scores for each question type. This data constitutes the multidimensional input of S14.

[0103] S14 uses a fuzzy comprehensive evaluation algorithm to synthesize multi-dimensional inputs into a single value for evaluating training effectiveness. , where B is the output comprehensive evaluation vector, W is the weight vector of each evaluation index, and R is the fuzzy relationship matrix between evaluation factors and evaluation levels.

[0104] Taking employee 001 as an example, four evaluation indicators are selected, namely, the learning time target completion rate. Course completion rate Interactive quiz accuracy and accompanying exam results Weight vectors of each indicator Among them, exam scores have the highest weight, while learning behavior indicators have a lower weight. The evaluation levels are divided into four levels: "Excellent," "Good," "Pass," and "Fail," with corresponding evaluation level vectors. .

[0105] The actual values ​​of the four indicators for employee 001 are: learning time completion rate Course completion rate Interactive quiz accuracy Supporting exam results These four actual values ​​are mapped to corresponding row vectors of the fuzzy relation matrix R according to the membership functions of each level, and the sum of the components in each row is normalized to 1. After mapping the actual values ​​of employee 001, the fuzzy relation matrix R is:

[0106] Performed using a weighted average composite operator The calculation yields the comprehensive evaluation vector. The sum of components B is 1.000. According to the principle of maximum membership, the largest component in B, 0.510, corresponds to a "Good" grade, meaning employee 001's training effectiveness evaluation grade is "Good". The overall score is determined by... The calculation shows that (Keep two decimal places).

[0107] If the preset evaluation threshold is set to 0.70, then the evaluation result of employee 001's training effectiveness will be... If the criteria for reaching the preset evaluation threshold are met, S1 grants employee 001 the authority to start work.

[0108] In some embodiments, the evaluation indicators for training effectiveness include training completion rate, examination pass rate, practical compliance rate, and accident rate. Training completion rate refers to the ratio of the number of course items completed by an employee to the number of course items that should be completed; examination pass rate refers to the ratio of the number of times an employee passes a corresponding examination within a statistical period to the total number of times they participate; practical compliance rate is driven by statistical data returned by the S2 violation monitoring subsystem and refers to the ratio of the number of days an employee did not violate regulations during their employment to the total number of days on duty; accident rate refers to the ratio of the number of safety accidents attributable to individual operation that occurred to an employee during their employment to the average level for their position. Of these four indicators, practical compliance rate and accident rate need to be backfilled from the S2 operational results. That is, after an employee passes the S14 assessment for the first time, data for these two indicators is only available during subsequent assessments. Therefore, these two indicators are mainly used for periodic retraining assessments of employees, rather than initial onboarding assessments.

[0109] In practical applications, preprocessing methods other than S11 to S14 can also achieve the training release objective of S1. For example, the implementation of S1 can also adopt the traditional process of "offline sign-in training + paper-based examination + manual grading" as an alternative path. In this case, the document parsing in S11 is replaced by manual entry of the employment form, the course generation in S13 is replaced by a general course pre-written by teaching and research personnel, and the multi-dimensional evaluation in S14 is replaced by a single score of the accompanying examination. This alternative path can also achieve the purpose of generating training content according to the job, evaluating the effect according to the employee's completion, and granting job access. Therefore, it is also a implementation method of S1. However, the embodiments of this application preferably use a combination of automated sub-steps of S11-S14 based on artificial intelligence technology to improve personalization and response speed.

[0110] S2. In response to an employee having the authority to work and entering the on-duty working state, perform violation identification on the employee's work video, obtain the violation judgment result, and update the employee's violation points based on the violation judgment result.

[0111] S2 is the on-duty supervision step of this method, triggered by two prerequisites: first, the employee has obtained work authorization through S1; second, the employee has actually entered the on-duty work state. The determination of on-duty work status can be achieved in various ways, such as the employee clocking in at the branch, the employee activating the delivery vehicle's onboard terminal, or location data from the employee's handheld device indicating that the employee has entered the delivery area. Taking employee 001 as an example, when employee 001 clocks in at the Chengdong branch and activates the delivery vehicle's onboard terminal, employee 001 is determined to have entered the on-duty work state, triggering the execution of S2.

[0112] The input-output closed loop of S2 is as follows: input is the work video of employee 001, output is the violation judgment result, and update employee 001's violation points according to the violation judgment result. This involves multiple sub-steps, including video preprocessing, violation recognition model inference, and violation judgment rule verification. Employee 001's violation point account is maintained independently by the system for each employee. The initial account value can be set to 12 points. When a violation occurs, the account value is deducted according to the different update increments corresponding to the violation level.

[0113] Specifically, refer to Figure 3 S2 includes sub-steps S21-S24.

[0114] S21. Acquire operational videos using visual acquisition devices, including fixed visual acquisition devices deployed in the operational area and vehicle-mounted visual acquisition devices deployed in delivery vehicles.

[0115] The deployment of visual acquisition equipment adopts a dual-track spatial division of "fixed + vehicle-mounted". Fixed visual acquisition equipment is deployed at key locations in the work area, such as the sorting table, security checkpoint, and entrances / exits of the Chengdong branch, covering employee 001's pickup and loading processes within the branch; vehicle-mounted visual acquisition equipment is deployed on the delivery electric vehicle driven by employee 001, covering employee 001's driving behavior and road environment during delivery. The two types of visual acquisition equipment work in parallel, collecting operational videos of different spatial areas without overlap or omission.

[0116] Taking employee 001 as an example, the fixed visual acquisition device has a frame rate of no less than 25fps, such as 30fps, an image resolution of no less than 1920×1080, supports 360° panoramic shooting, and has infrared night vision function to cover nighttime operation scenarios; the vehicle-mounted visual acquisition device has the same frame rate and resolution as the fixed visual acquisition device, but the field of view is focused on the road in front of the delivery vehicle and the employee's driving posture.

[0117] S22. Perform preprocessing on the work video to obtain the preprocessed work video.

[0118] The purpose of preprocessing is to eliminate factors in the work video that are detrimental to the inference of the subsequent violation detection model, such as noise, lighting deviation, and resolution redundancy, so that the input of the subsequent violation detection model has uniform data characteristics. S22 performs preprocessing on the work video of employee 001. The specific preprocessing operations include denoising, cropping, and scaling. Denoising uses a Gaussian filter to convolve the video frames to reduce sensor noise and compression noise; cropping is to extract regions of the video frames according to the effective field of view of the fixed vision acquisition device or the vehicle-mounted vision acquisition device, discarding black borders and irrelevant backgrounds; scaling normalizes the resolution of the video frames to the input size of the violation detection model (e.g., 640×640).

[0119] For example, the original video frame of employee 001 working at the sorting station in the Chengdong branch has a resolution of 1920×1080, a frame rate of 30fps, and contains slight sensor noise. After S22 preprocessing, the video frame is normalized to a resolution of 640×640, a frame rate of 25fps, and the noise is significantly reduced. This preprocessed video can be directly used as input for the subsequent violation identification model.

[0120] S23. Input the preprocessed operation video into the violation identification model. The violation identification model is trained using a violation sample set from the express delivery industry. The violation identification model performs feature extraction and classification on the preprocessed operation video and outputs information on suspected violations.

[0121] The violation detection model is a target detection model custom-trained for violation scenarios in the express delivery industry. Its input is a single or multi-frame tensor of preprocessed operational videos, and its output includes the location region of the suspected violation, the violation category, and the confidence score. The training dataset for the violation detection model is a violation sample set from the express delivery industry, including employee operational violations such as rough handling, not wearing gloves, not wearing helmets, and haphazardly stacking packages; and traffic violations during delivery such as running red lights, driving against traffic, not driving in designated lanes, fatigued driving, and using a mobile phone while driving. During the training phase, the violation detection model learns the visual features of various violations, and during the inference phase, it extracts features and classifies the preprocessed operational videos of employee 001.

[0122] Taking a video frame from a delivery route of employee 001 as an example, the vehicle-mounted visual acquisition device captured employee 001's delivery electric vehicle entering a one-way street. After preprocessing by S22, this video frame is input into the violation identification model. The violation identification model outputs the following suspected violation information: violation category "driving against the flow of traffic", location area of ​​the violation (timestamp of the corresponding video frame and coordinates of the area in the frame), and confidence level 0.99.

[0123] In some embodiments, the express delivery industry violation sample set includes employee operational violation samples and delivery traffic violation samples. This sample set undergoes data augmentation processing to expand the sample size before being input into the violation recognition model for training. Data augmentation operations include performing random rotation, random cropping, random scaling, random brightness adjustment, and noise addition on the sample images, generating multiple enhanced samples with slightly different visual features but unchanged violation categories for each original sample. For example, an original sample of "delivery person going against traffic" can be expanded into 10 to 20 enhanced samples after data augmentation, covering visual variations of day / dusk / night, different weather conditions, and different road backgrounds. This enables the violation recognition model to have a high generalization ability for unfamiliar real-world scenarios after training.

[0124] S24. Verify suspected violation information using preset violation judgment rules, and output violation judgment results, including violation type and violation level.

[0125] Suspected violation information originates from the inference output of the violation identification model. While possessing high confidence, it is not yet aligned with business rules such as safety operating procedures and traffic regulations within the express delivery industry. S24 performs secondary verification on the suspected violation information using preset violation judgment rules, matching it against rule entries in the business rule base to output the final violation judgment result. The preset violation judgment rules must at least include a mapping relationship of "suspected behavior category → violation type → violation level," as well as relevant business constraints (such as speed limits in delivery areas, one-way street attributes for specific road sections, etc.).

[0126] Taking employee 001's suspected violation of driving against traffic as an example, the corresponding rule entry in the preset violation judgment rules is: suspected behavior category "driving against traffic" occurring on a one-way road section → violation type "traffic violation - driving against traffic" → violation level "major violation". After S24 matches the suspected violation information with the rule entry, it outputs the violation judgment result: violation type is "traffic violation - driving against traffic", violation level is "major violation", violation occurrence time is the timestamp corresponding to the vehicle-mounted visual acquisition device, and violation occurrence location is the GPS latitude and longitude coordinates. This violation judgment result will be passed to the subsequent graded response operation and violation point update stage.

[0127] In some embodiments, violation levels are divided into general violations, serious violations, and major violations, with different update ranges for violation points corresponding to different violation levels. For example, a general violation corresponds to a deduction of 1 to 3 points, a serious violation to a deduction of 4 to 6 points, and a major violation to a deduction of 12 points. Violation points are recorded as cumulative deductions. Employee 001's violation point account has a cumulative deduction of 0 points upon initialization. After each violation judgment result is output, the deduction value is added to the cumulative deduction value according to the update range corresponding to the violation level. This reverse driving violation is judged as a major violation, and the update range for violation points is 12 points. After the update, employee 001's cumulative violation point value is 12 points. The cumulative deduction value will serve as the direct basis for S3 to determine whether to trigger intensive training.

[0128] When multiple suspected violations are identified concurrently within the same time window, the S24's preset violation judgment rules process them according to the priority of the violation level from high to low. The violation judgment result with the highest level is output first and drives the subsequent process, while the other lower-level violation judgment results are processed sequentially to avoid omissions.

[0129] In practical applications, implementations of violation identification other than S21 to S24 can also support the on-duty supervision purpose of S2. For example, the implementation of S2 can also be based on a combination of audio recognition and location trajectory: the visual acquisition in S21 can be replaced by a microphone array to collect sound from the employee's work site, and the visual model inference in S23 can be replaced by an audio classification model to identify abnormal acoustic features (such as impact sounds from violent sorting, or sounds of arguments), and perform spatiotemporal matching with the location trajectory of the employee's handheld terminal. This alternative path can also achieve the purpose of identifying violations from the signal input of the employee's work and generating violation judgment results, and therefore also belongs to the implementation of S2.

[0130] In some embodiments, the violation detection model is built on an object detection network. A channel attention structure is embedded in the feature extraction layer of the object detection network. The channel attention structure is used to perform weight recalibration on the feature channels to obtain enhanced features for violations of small targets.

[0131] The object detection network for the violation detection model can be the YOLO series or other single-stage object detection architectures. This application embodiment does not limit the specific object detection network selection. A typical object detection network structure consists of three parts: a backbone feature extraction layer, a neck feature fusion layer, and a detection head. The backbone feature extraction layer is responsible for extracting multi-scale feature maps from the preprocessed input video. This application embodiment embeds a channel attention structure in the backbone feature extraction layer to recalibrate the weights of the feature channels output by the backbone feature extraction layer, highlighting channels that contribute significantly to violation detection and suppressing channels that are irrelevant or redundant to violation detection.

[0132] The core idea of ​​the channel attention structure is a two-step operation of "compression-activation". The compression step performs global average pooling on each channel of the feature map, dividing each channel into... The spatial resolution is compressed into a scalar to obtain the channel descriptor vector z; the activation step feeds the channel descriptor vector z into a small two-layer fully connected network, outputting the weight coefficient s of each channel. The weight coefficients are normalized to the numerical range between 0 and 1 by the Sigmoid function; finally, the original feature map of each channel is multiplied by the corresponding weight coefficient by channel to obtain the recalibrated feature map.

[0133] The calculation for the compression step is as follows: ,in This is the original feature map of the c-th channel. Let H be the descriptor scalar for the c-th channel, and H and W be the height and width of the feature map, respectively. The calculation of the activation step is as follows: ,in and The weight matrix is ​​for a two-layer fully connected system. It is the ReLU activation function. This is the Sigmoid activation function.

[0134] Taking a feature map with 256 channels as an example, the channel descriptor vector z output by the compression step is a 256-dimensional vector, and the weight coefficient vector s output by the excitation step is also a 256-dimensional vector. For the delivery video frame of employee 001, the feature channels with a high correlation to violations, such as those corresponding to traffic light colors and vehicle driving directions, have higher weight coefficients calculated by the excitation step; while the channels that are unrelated to violations or are redundant, such as those corresponding to background roadside trees and distant buildings, have lower weight coefficients calculated by the excitation step. In the recalibrated feature map, features highly correlated with violations such as "traffic light colors" and "vehicle driving directions," which are related to "driving against traffic" or "running a red light," are significantly enhanced, while redundant background features are suppressed. This enables the detection head to have a high detection rate for violations of small targets, such as the red light status of distant traffic lights and arrows indicating the direction of illegal driving.

[0135] For violations that occupy a small pixel area in the image, such as not wearing gloves or improperly placing packages, the recalibration mechanism of the channel attention structure increases the weight of the relevant feature channels output by the feature extraction layer. Small-target violations that were originally easily submerged by background features due to their small pixel proportion have a higher signal-to-noise ratio in the recalibrated feature map, making the classification results of the detection head more reliable.

[0136] In some embodiments, the weights of the violation detection model adopt the INT8 quantization format, which is suitable for deployment on embedded devices and vehicle terminals. INT8 quantization maps the weight parameters in the violation detection model, which were originally represented as 32-bit floating-point numbers, to 8-bit integers. The model file size is compressed to about one-quarter of the original size, and the computing power requirements and power consumption during inference are also significantly reduced. This allows the violation detection model to be deployed on the vehicle terminal of the delivery vehicle driven by employee 001 or on the edge computing device of the Chengdong branch. In scenarios where the network is interrupted or bandwidth is limited, the violation detection model can complete inference and output violation judgment results locally, so that S2's violation detection does not depend on the real-time availability of cloud computing power and can run continuously.

[0137] Furthermore, the channel attention structure of the violation detection model is not the only feature recalibration scheme. In some embodiments, a spatial attention structure can be used to recalibrate the spatial location dimension of the feature map, making the model more sensitive to the spatial region where the violation occurs; or a hybrid attention structure can be used to perform recalibration of both the channel and spatial dimensions simultaneously. These alternative paths can also achieve feature enhancement for small target violations, and therefore also belong to the implementation methods of the violation detection model.

[0138] After S2 obtains the violation determination result, the following graded response operations are also included.

[0139] The tiered response system uses a preset trigger threshold as a boundary to divide violation responses into two levels: minor and severe. The preset trigger threshold is aligned with the violation level: minor violations do not reach the preset trigger threshold, while serious and major violations do. For employee 001, different violation scenarios will trigger different response branches.

[0140] In response to a violation determination result indicating that the severity of the violation has not reached a preset threshold, a reminder is issued to the employee via their handheld terminal and on-site voice broadcast. For example, if employee 001 is identified as "not wearing gloves" during sorting, this violation is classified as a minor violation, and its severity has not reached the preset threshold. At this time, a reminder pop-up window appears on employee 001's handheld terminal: "Please note that you have violated the rules; please correct them immediately." Simultaneously, the broadcast system in the sorting area of ​​the Chengdong branch broadcasts the same message. The violation identification model outputs the violation determination result within 300ms. The determination result is transmitted to employee 001's handheld terminal via the reminder link. The end-to-end response latency of the reminder pop-up window is controlled within 1 second, and the trigger latency of the voice broadcast is controlled within 2 seconds, ensuring that employee 001 receives feedback and corrects the violation immediately during its duration.

[0141] In response to a violation determination indicating that the severity of the violation has reached a preset threshold, the system not only alerts the employee but also simultaneously sends an intervention alert to the management terminal and a device linkage command to the equipment control system, triggering the braking of the corresponding equipment. For example, if employee 001 is identified as "driving against the flow of traffic" during a delivery, this violation is classified as a major violation, reaching the preset threshold. In addition to alerts issued by employee 001's handheld terminal and the vehicle's voice broadcast, an intervention alert pop-up appears on the administrator's terminal at the Chengdong branch, containing fields such as the violating employee's ID, violation type, violation location, and a link to a video clip related to the violation, facilitating rapid follow-up by the administrator. Simultaneously, the vehicle's control system receives the device linkage command, applies speed-limited braking to the delivery vehicle, limiting the speed to no more than 15 km / h, and triggers an audible and visual alarm on the delivery vehicle, making employee 001 aware that the equipment has entered a restricted state and cannot quickly escape the violation on its own, requiring immediate stopping and correction.

[0142] The violation determination result, the corresponding work video clip, and the identification information indicating the employee's identity are encapsulated into a violation record entry, which is then written to the violation record storage area. A violation record entry is a structured data record whose fields include at least all fields of the violation determination result, the work video clip (a segment extracted from employee 001's work video before and after the violation occurred, e.g., a 10-second clip from 5 seconds before to 5 seconds after the violation), and identification information indicating the employee's identity (such as employee ID E001, employee name, and the employee's branch office, Chengdong branch, etc.). The violation record storage area is a dedicated partition in the data storage layer for storing violation record entries. It supports multi-dimensional queries by employee, position, time, and violation type, providing a traceable data source for subsequent performance evaluation, accountability, points updates, and iterative optimization of the violation identification model.

[0143] In some embodiments, violation records are stored using blockchain technology. A core feature of blockchain storage is that each violation record, after being written, generates a hash value associated with the previous record and is simultaneously broadcast to multiple nodes, ensuring that once a violation record is entered into the chain, it cannot be tampered with. After employee 001's reverse violation record is entered into the chain, its violation determination result, video clip hash, timestamp, and other fields are all saved in an immutable form. This provides a verifiable chain of evidence for employees to appeal when they have doubts about the violation determination, and also provides credible traceability for enterprises facing compliance audits by industry regulatory authorities.

[0144] S3. In response to the violation points meeting the conditions for triggering enhanced training, an enhanced training instruction is generated and sent back to S1. The enhanced training instruction is used to drive the regeneration and re-execution of training content for employees, and locks the employee's work permissions during the regeneration and re-execution. This allows the violation judgment result to be applied to the regeneration of training content through the enhanced training instruction, thus constructing a feedback link from S2 to S1.

[0145] S3 is the feedback loop link of this method, responsible for feeding the results of the violation monitoring link back to the training link. Employee 001's violation points are updated in real time after each violation judgment result output by S2, and S3 determines whether employee 001's violation points meet the conditions for triggering reinforcement training after each update.

[0146] In some embodiments, the triggering conditions for enhanced training include at least one of the following: the cumulative value of violation points within a preset statistical period reaches a preset point threshold; the violation level corresponding to the violation points reaches a preset level; or the violation type corresponding to the violation points is a preset high-risk type. These three triggering conditions complement each other, enabling enhanced training to be triggered for different violation patterns.

[0147] For the first trigger condition, taking a preset statistical period of 30 days and a preset points threshold of 12 points as an example, if employee 001 is judged as a general violation 4 times within 30 days, with each violation increasing by 3 points, the cumulative deduction value is 12 points. When the cumulative deduction value reaches the preset points threshold, the trigger condition is activated.

[0148] For the second type of triggering condition, taking the preset level of major violation as an example, if employee 001 is judged to have committed a major violation in a certain violation judgment result (such as the reverse driving violation in this case), the triggering condition will be activated immediately without the need to accumulate points to the threshold.

[0149] For the third trigger condition, taking "fatigue driving" as the preset high-risk type as an example, if employee 001 is identified as having a fatigue driving violation, even if the penalty for violation points has not yet reached the preset point threshold according to the serious violation level (4 to 6 points), the trigger condition will be activated because the violation type matches the preset high-risk type.

[0150] Taking employee 001's violation of driving against traffic in the above scenario as an example, the violation judgment result corresponds to a major violation, which meets the second triggering condition. After S3 confirms that employee 001's violation points meet the triggering condition for enhanced training, it generates an enhanced training instruction. The enhanced training instruction is a structured instruction message, with fields including employee ID E001, triggering reason "major violation - driving against traffic", historical violation summary (summary of employee 001's violation records in the past 30 days), and instruction generation timestamp, etc.

[0151] After the reinforcement training instruction is generated, it is transmitted back to S1 via the internal signal path of the system. During the transmission, the fields of the reinforcement training instruction are parsed by the receiving end of S1: employee ID E001 is used to locate the structured employee file of employee 001, the triggering reason "major violation - wrong-way driving" is used to guide the course generation model of S13 to generate new training content including "wrong-way driving special training" and "traffic safety enhancement", and the historical violation summary is used as an auxiliary input for the course generation model of S13, so that the new training content can be arranged to target the specific weaknesses of employee 001.

[0152] At the same moment the enhanced training instruction reaches S1, employee 001's work permissions are locked. While work permissions are locked, employee 001's clocking in and starting of the delivery vehicle cannot trigger the execution of S2, and employee 001 is placed in an "unable to work" state at the system level. The work permission lock state continues until employee 001 completes the regenerated training content, passes the regenerated accompanying exam, and the training effectiveness evaluation result in S14 again reaches the preset evaluation threshold. After passing the re-evaluation, employee 001's work permissions are unlocked, and employee 001 can resume on-duty operations.

[0153] Through the aforementioned closed loop, the violation judgment result of S2 serves as input. Through the accumulation of violation points and the trigger condition judgment of S3, it ultimately drives S1 to re-execute the training release process for the same employee, forming an automated feedback loop from the violation monitoring stage to the training stage. This feedback loop ensures that employee 001's on-the-job performance data after initially passing S1 is no longer decoupled from the training content. Each violation by employee 001 will cumulatively affect their eligibility for the position. Employees with repeated violations will be automatically pushed into intensive training and continuously corrected through the feedback loop until employee 001's violations are significantly reduced or eliminated.

[0154] Furthermore, the method also includes the S4 emergency query stage and the S5 weather warning stage. The module division and interaction relationship of the two stages are shown in Figure 4.

[0155] S4. Obtain emergency query texts from employees, perform semantic parsing on the emergency query texts using a large-scale safety model, and combine the emergency database to output corresponding handling plan suggestions. The emergency database includes a historical emergency event experience database and an emergency plan database.

[0156] S4 is designed to provide real-time emergency guidance to employees facing unexpected events while on duty. When employee 001 encounters an emergency while delivering packages, such as witnessing a passerby suddenly fall ill, being involved in a traffic accident, or a package emitting smoke, they can open the emergency interaction interface on their handheld terminal and input emergency query text in natural language. The emergency query text can be a short phrase, such as "The package is smoking," or a complete sentence, such as "What should I do if I encounter a pedestrian suddenly fainting while delivering a package?" If employee 001's handheld terminal supports voice input, their voice will be processed into speech-to-text and then transmitted to S4 as the emergency query text.

[0157] After receiving the emergency query text, S4 performs semantic parsing on it using the safety big data model. The output of the semantic parsing is a structured semantic representation, which includes at least the emergency event type (e.g., "operational accident - package spontaneous combustion," "sudden illness," or "traffic accident"), key entities of the emergency event (e.g., the items involved, the personnel involved, or the location), and the intent of the emergency query (e.g., requesting assistance with procedures, contact information, or on-site first aid methods). Based on the semantic representation, the safety big data model retrieves relevant historical emergency events and emergency plans from the emergency database and generates suggested solutions corresponding to the emergency query text based on the search results. The emergency database includes two sub-databases: a historical emergency event experience database and an emergency plan database. The historical emergency event experience database stores records of past emergency events handled in the express delivery industry, with fields including event type, occurrence scenario, handling process, handling effect, and experience summary. The emergency plan database stores standardized emergency plans for various emergency scenarios, with fields including applicable scenarios, handling procedures, division of responsibilities, precautions, and material allocation plans.

[0158] Taking the scenario where employee 001 encounters a "package emitting smoke" while delivering a package as an example, employee 001 enters the emergency query text "What to do if a package is emitting smoke while delivering a package" on the emergency interaction interface of a handheld terminal. The semantic analysis result of the safety big data model for this text is: emergency event type "operational accident - package spontaneous combustion", key entity "package", emergency query intent "help with handling steps". Based on this semantic representation, the safety model retrieved three past "package spontaneous combustion" incident records from the historical emergency event experience database and one "package spontaneous combustion" emergency plan from the emergency plan database. Combining the search results, it generated the following suggested response: "First, immediately evacuate surrounding personnel; do not attempt to extinguish the fire with water (it may be a lithium battery); second, use a dry powder fire extinguisher to spray from the side of the package; third, call 119 and notify the branch supervisor. Note: If the fire is larger than a washbasin, evacuate immediately and do not attempt to handle it yourself." This suggested response was returned to employee 001's handheld terminal within one second of the employee sending the emergency query text. Employee 001 then executed the on-site response actions based on the suggested response.

[0159] In some embodiments, the safety big data model is built on the Transformer architecture and pre-trained and fine-tuned using an emergency database. During the pre-training phase, the safety big data model learns basic language understanding and generation capabilities on a general text corpus. During the fine-tuning phase, the safety big data model undergoes supervised fine-tuning on "emergency query text → suggested handling plan" sample pairs constructed from emergency database data, enabling it to possess specialized semantic understanding capabilities for emergency scenarios in the express delivery industry. The fine-tuned safety big data model demonstrates significantly higher accuracy in responding to express delivery industry-specific emergency scenarios such as "package spontaneous combustion," "delivery personnel sudden illness," and "sorting equipment failure" compared to the un-fine-tuned general big data model.

[0160] In some embodiments, in response to the safety big data model's parsing result of the emergency query text not matching a valid item in the emergency database, a prompt for manual emergency command access is generated, and the applicable scenario and capability boundaries are marked in the output disposal plan suggestion. The determination of a mismatch can be made in two ways: first, the highest similarity of the emergency database search result is lower than a preset similarity threshold (e.g., 0.5); second, the confidence assessment output within the safety big data model is lower than a preset confidence threshold. In the case of a mismatch, the safety big data model does not forcibly generate a disposal plan suggestion, but instead outputs a prompt for manual emergency command access. The emergency interaction interface on employee 001's handheld terminal displays "The current emergency scenario exceeds the scope of automatic emergency guidance capabilities. Please immediately call the manual emergency command number XXX," and directly presents a button for one-click manual emergency command. Even in the case of a match, the disposal plan suggestion text will be appended with markings of applicable scenarios and capability boundaries, such as "This solution is applicable to routine express delivery operations. If hazardous materials or known lithium battery fires are involved, please contact the fire department immediately," so that employee 001 clearly understands the limitations of the safety big data model's output.

[0161] In some embodiments, completed emergency response records are populated back into the historical emergency event experience database for incremental training of the safety model. After handling a "package spontaneous combustion" incident, employee 001 can fill in fields such as the actual handling process, handling effect, and experience summary on the emergency interaction interface. After submission, this new emergency response record is populated back into the historical emergency event experience database. When the historical emergency event experience database accumulates a certain number of new records, the safety model performs a round of incremental training, incorporating the new records into the training samples and fine-tuning the model parameters. After incremental training is completed, the safety model's response accuracy to similar emergency events is further improved, and the emergency response capability continues to evolve as the method's runtime increases.

[0162] Furthermore, this AI-based method for personnel safety management in the express delivery industry also includes: S5. Obtain meteorological risk data for the work area, determine the warning level based on the meteorological risk data, and generate warning push messages for employees based on the warning level.

[0163] S5 is responsible for issuing early warnings of meteorological risks in the work area before or during employee work. Meteorological risk data comes from real-time data streams connected to a public interface with meteorological departments. Fields include meteorological disaster type (e.g., typhoon, rainstorm, blizzard, high temperature, cold wave), meteorological warning level, affected area, and duration. S5 continuously monitors meteorological risk data for the Chengdong area where employee 001 is located. When meteorological risk data indicates a meteorological risk in the area, S5 determines the warning level based on the data and generates a warning push notification for employee 001. The warning push notification includes fields such as warning level, warning type, affected area, duration, and response plan, and is simultaneously pushed to employee 001 and the branch administrator via their handheld terminal and the display screen at the Chengdong branch.

[0164] In some embodiments, the determination of the warning level also refers to the geographic environment data of the work area. The geographic environment data is maintained by a geographic information system (GIS), and fields include the terrain features of the work area (e.g., urban areas, suburbs, mountainous areas), road condition features (e.g., main roads, secondary roads, inner streets), and the density of distribution points. Warning levels for different work areas under the same meteorological risk level can be differentiated. For example, taking employee 001's eastern urban area and a mountainous delivery area as examples, both receive a yellow rainstorm warning. Based solely on meteorological risk data, their warning levels are the same; however, after overlaying geographic environment data, the eastern urban area, due to its flat terrain and good road conditions, maintains a yellow warning level; while the mountainous delivery area, due to its undulating terrain and the risk of flash floods in some sections, has its warning level upgraded to a red warning. This differentiated determination mechanism makes the warning push information generated by S5 more closely aligned with the actual risks of specific work areas.

[0165] In some embodiments, the warning levels are divided into red, yellow, and blue warnings, with different warning levels corresponding to different outdoor operation restriction policies. A red warning corresponds to "prohibiting all outdoor delivery operations and ceasing all outdoor operations at the sorting center"; a yellow warning corresponds to "restricting outdoor delivery operations, shortening outdoor operation time, and suspending high-risk operations such as high-altitude operations"; and a blue warning corresponds to "reminding employees to take appropriate precautions against current weather risks, such as wearing warm clothing in cold weather and rain gear in heavy rain, and paying attention to operational safety." Employee 001 and the Chengdong branch administrator can directly execute the corresponding operational adjustments based on the received warning level. For example, if employee 001 returns to the Chengdong branch immediately after receiving a red warning, the administrator will postpone the delivery tasks for the Chengdong area until the warning is lifted.

[0166] In some embodiments, monitoring strategy adjustment parameters are generated based on the warning level and injected into the violation identification of S2. This causes the identification focus and violation judgment rules of S2 to adjust synchronously with the warning level, enabling S5 and S2 to form a cross-linkage, supplementing the monitoring-to-training feedback link constructed by S3 as another cross-linkage link. The injection of monitoring strategy adjustment parameters into S2 occurs after S2 has become operational. When S2 has been continuously running for employee 001, changes in the warning level of S5 trigger the generation and injection of monitoring strategy adjustment parameters, causing S2 to operate according to the new identification focus and violation judgment rules in subsequent violation identification.

[0167] The specific forms of parameter adjustment for the supervision strategy include weight adjustment for identifying key points and bias adjustment for violation judgment rules. Weight adjustment for identifying key points affects the reasoning process of the violation identification model. For example, when the warning level is red, the behavior of "outdoor delivery" is itself a violation, and the violation identification model increases the identification weight of the "outdoor delivery" type of violation. Bias adjustment for violation judgment rules affects the rule matching process of S24. For example, when the warning level is yellow, "delivering packages for an extended period of time in the rain," which was originally judged as a "serious violation," is biased and adjusted to "major violation" to trigger a more severe graded response. Taking a delivery day after employee 001 completed intensive training and regained their work privileges as an example, at 10:00 AM that day, S5 received a red rainstorm warning, and the monitoring strategy adjustment parameters were immediately generated and injected into S2; employee 001 was identified by the vehicle-mounted visual acquisition device as delivering packages outdoors, and S2's violation identification model judged the behavior as "outdoor delivery - during the red rainstorm warning period" based on the adjusted identification weight, and the violation judgment result was "major violation", triggering a graded response, including vehicle braking and administrator intervention reminders, urging employee 001 to immediately return to the Chengdong branch.

[0168] This application also provides an artificial intelligence-based personnel safety management system for the express delivery industry, referring to... Figure 1 The system includes a personnel intelligent training subsystem, a personnel violation supervision subsystem, and an interface adaptation layer that connects the personnel intelligent training subsystem and the personnel violation supervision subsystem.

[0169] The intelligent personnel training subsystem is configured to execute the function corresponding to method step S1. Specifically, it performs the following actions: generating training content for employees based on their job-related data; generating training effectiveness evaluation results based on employees' completion of the training content; and granting employees job access upon reaching a preset evaluation threshold. The intelligent personnel training subsystem can be further divided into an identity resolution module, a job characteristic resolution module, a course generation module, a supporting examination module, a training tracking module, and an effectiveness evaluation module, each responsible for executing sub-steps S11 to S14.

[0170] The personnel violation monitoring subsystem is configured to execute functions corresponding to method step S2 and the tiered response operations following S2. Specifically, it performs the following actions: acquiring work videos of employees with work authorization and in an on-duty work state; performing violation identification on the work videos to obtain violation judgment results; updating the employee's violation points based on the violation judgment results; and generating a reinforcement training instruction when the violation points meet the trigger conditions for reinforcement training. The personnel violation monitoring subsystem can be further divided into components such as visual acquisition devices, a preprocessing module, a violation identification model, a violation judgment module, a real-time alert module, a device linkage module, a violation record storage area, and a violation point account, each responsible for executing steps S21 to S24 and the tiered response operations, respectively.

[0171] The interface adaptation layer is configured to carry out the feedback link function of S3. Specifically, it transmits reinforcement training instructions from the personnel violation monitoring subsystem to the personnel intelligent training subsystem, drives the regeneration and re-execution of training content for employees, and locks employees' work permissions during the regeneration and re-execution process. As the physical carrier of the feedback link, the interface adaptation layer ensures that reinforcement training instructions have a unified message format, unified routing rules, and a unified secure transmission mechanism during cross-subsystem transmission.

[0172] In some embodiments, the system further includes a personnel safety emergency response subsystem. This subsystem is configured to acquire meteorological risk data for the work area, determine the warning level based on the meteorological risk data, generate warning push notifications for employees based on the warning level, and simultaneously handle semantic parsing of emergency query text and generation of suggested response plans. The personnel safety emergency response subsystem corresponds to steps S4 and S5 of the method and can be internally divided into components such as a text interaction module, a safety big data model, an emergency database, a meteorological risk interface, a warning classification module, and a warning push module.

[0173] In some embodiments, the system further includes a data storage layer and an algorithm support layer. (Continue to refer to...) Figure 1The data storage layer is configured to store job-related data, training effectiveness evaluation results, violation judgment results, and violation points. The algorithm support layer is configured to support the operation and updates of the violation identification model. The data storage layer adopts a distributed storage architecture combining relational and non-relational databases. The relational database stores structured data such as employee information, training data, violation records, and point data, while the non-relational database stores unstructured data such as video data, image data, emergency event data, and emergency plan documents. The data storage layer supports data sharding and disaster recovery backup. The relational database uses a relational database management system as its underlying layer, suitable for storing data types with fixed field structures and requiring transaction consistency guarantees. The non-relational database uses a document-oriented or key-value database management system as its underlying layer, suitable for storing unstructured data with flexible fields and large volumes. Data sharding partitions large tables across multiple physical nodes according to employee ID or time range. Disaster recovery backup maintains copies of each data on multiple nodes, ensuring that other nodes can still provide complete service even if one node fails. The algorithm support layer is built on a deep learning framework, computer vision algorithm library, and natural language processing algorithm library, supporting online updates and iterations of algorithm models. The algorithm support layer can also deploy an algorithm model monitoring module to monitor inference latency, recognition accuracy, and other indicators of the violation identification model, course generation model, and security big model in real time. When the indicators drop, an alarm is triggered and the model rollback or retraining process is started.

[0174] In some embodiments, the interface adaptation layer includes device interfaces, third-party interfaces, and internal interfaces. Device interfaces are used to interface with hardware devices, such as visual acquisition devices, operational equipment control systems, employee handheld terminals, and branch network displays. Third-party interfaces are used to interface with external system interfaces, such as publicly available interfaces from meteorological departments, interfaces from existing enterprise management systems, and identity authentication interfaces. Internal interfaces are used to enable data interoperability between subsystems, such as the transmission of enhanced training instructions between the personnel violation monitoring subsystem and the personnel intelligent training subsystem, and the transmission of monitoring strategy adjustment parameters between the personnel safety emergency subsystem and the personnel violation monitoring subsystem. The interface adaptation layer uses the RESTful protocol as a unified communication protocol and performs encryption processing on data transmission, such as using symmetric encryption algorithms to encrypt the message body, ensuring both general compatibility and security during cross-interface data transmission.

[0175] In some embodiments, the system deployment adopts a collaborative deployment mode between the cloud and the edge. The cloud deploys the backend management, security big data model, and data storage layer, while the edge deploys the violation identification algorithm, visual acquisition devices, and device linkage. The edge supports offline operation, and offline data is synchronized to the cloud after network recovery. The cloud carries computationally intensive and data-intensive functional modules, utilizing the elastic computing power of public or enterprise private clouds to support the storage of large-scale employee data and the inference of large models. The edge carries functional modules with high real-time requirements, such as S2's violation identification, which must output violation judgment results within 300ms after an employee's violation occurs. Such low-latency requirements are not suitable for relying on round-trip transmission from the cloud. The offline operation capability of the edge addresses scenarios with widely distributed express delivery outlets and variable network environments. Taking the Chengdong outlet as an example, when the network link of the Chengdong outlet is interrupted, the violation identification algorithm deployed on the edge computing device of the Chengdong outlet can still continuously identify the violation behavior of employee 001. The locally generated violation record entries are temporarily stored in the edge cache area and automatically synchronized to the violation record storage area in the cloud when the network link is restored.

[0176] In some embodiments, the system also includes a back-end management subsystem. This subsystem is configured to provide employee management, training management, violation management, emergency management, system configuration, and data statistical analysis functions. It also supports the generation of multi-dimensional safety management statistical reports by time, position, and region. The back-end management subsystem provides a unified control entry point for administrators. Administrators can view data such as employee 001's training progress, violation records, points balance, and emergency response records, and perform operations such as adding / editing / deleting employees, adjusting training courses, handling violation appeals, resetting violation points, updating the emergency database, and adjusting warning thresholds. The data statistical analysis function supports multi-dimensional statistics by time dimension (e.g., daily, weekly, monthly, quarterly, or year), by position dimension (e.g., sorter, delivery person, or security inspector), and by region dimension (e.g., various branches or areas). It generates statistical reports on indicators such as training pass rate, violation incidence rate, warning response rate, and accident incidence rate, providing a data view for enterprise safety management decisions.

[0177] Specific limitations regarding the AI-based personnel safety management system for the express delivery industry can be found in the above-mentioned limitations on AI-based personnel safety management methods for the express delivery industry, and will not be repeated here. The various modules in the aforementioned AI-based personnel safety management device for the express delivery industry can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0178] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database contains data related to an AI-based method for managing the safety of personnel in the express delivery industry. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an AI-based method for managing the safety of personnel in the express delivery industry.

[0179] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the artificial intelligence-based personnel safety management method for the express delivery industry described in the above embodiment.

[0180] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the artificial intelligence-based personnel safety management system for the express delivery industry described in the above embodiment.

[0181] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments of this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0182] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0183] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for personnel safety management in the express delivery industry based on artificial intelligence, characterized in that, Includes the following steps: S1. Generate training content for the employee based on the employee's job-related data, and generate a training effectiveness evaluation result based on the employee's completion of the training content. In response to the training effectiveness evaluation result reaching a preset evaluation threshold, grant the employee the right to take up the job. S2. In response to the employee having the work authorization and entering the on-duty work state, perform violation identification on the employee's work video, obtain the violation judgment result, and update the employee's violation points according to the violation judgment result; S3. In response to the violation points meeting the enhanced training trigger condition, an enhanced training instruction is generated and sent back to S1. The enhanced training instruction is used to drive the regeneration and re-execution of the training content for the employee, and during the regeneration and re-execution, the employee's work permission is locked, so that the violation judgment result is applied to the regeneration of the training content through the enhanced training instruction, thus constructing the feedback link from S2 to S1.

2. The artificial intelligence-based personnel safety management method for the express delivery industry according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Perform structured parsing on the employee's identification information and professional qualification information to obtain a structured employee file; S12. Match the job assignment information and safety risk points in the job-related data from the job characteristic database, and output a job training needs list; S13. Input the job training needs list and historical accident case data into the course generation model, output the training content, and generate a matching exam based on the training content and the historical accident case data that is compatible with the training content; S14. Collect data on the employee's learning time, course completion progress, and learning interaction with the training content, and combine this data with the results of the accompanying exam to generate the training effectiveness evaluation result using a fuzzy comprehensive evaluation algorithm.

3. The artificial intelligence-based personnel safety management method for the express delivery industry according to claim 1, characterized in that, S2 includes the following sub-steps: S21. The operation video is acquired by a visual acquisition device, which includes a fixed visual acquisition device deployed in the operation area and a vehicle-mounted visual acquisition device deployed in the delivery vehicle; S22. Perform preprocessing on the work video to obtain a preprocessed work video; S23. Input the preprocessed operation video into the violation identification model, which is trained using a violation sample set of the express delivery industry. The violation identification model performs feature extraction and classification identification on the preprocessed operation video and outputs suspected violation information. S24. Verify the suspected violation information in conjunction with preset violation judgment rules, and output the violation judgment result, which includes the violation type and violation level.

4. The artificial intelligence-based personnel safety management method for the express delivery industry according to claim 3, characterized in that: The violation identification model is built on a target detection network. A channel attention structure is embedded in the feature extraction layer of the target detection network. The channel attention structure is used to perform weight recalibration on the feature channels to obtain enhanced features for violations of small targets.

5. The artificial intelligence-based personnel safety management method for the express delivery industry according to claim 1, characterized in that, After obtaining the violation determination result in step S2, the following graded response operation is also included: If the severity of the violation indicated by the violation determination result does not reach the preset linkage threshold, a reminder will be issued to the employee via the employee's handheld terminal and on-site voice broadcast; In response to the violation determination result indicating that the severity of the violation has reached the preset linkage threshold, in addition to issuing a reminder to the employee, an intervention reminder is simultaneously issued to the management personnel terminal, and an equipment linkage command is sent to the work equipment control system to trigger the work equipment control system to brake the corresponding work equipment; The violation determination result, the corresponding work video clip, and the identification information indicating the employee's identity are encapsulated into a violation record entry, and the violation record entry is written into the violation record storage area.

6. The artificial intelligence-based personnel safety management method for the express delivery industry according to claim 1, characterized in that, It also includes the following steps: S4. Obtain the emergency query text from the employee, perform semantic parsing on the emergency query text using a large-scale security model, and output a suggested handling plan corresponding to the emergency query text in conjunction with the emergency database. The emergency database includes a historical emergency event experience database and an emergency plan database.

7. The artificial intelligence-based personnel safety management method for the express delivery industry according to claim 1, characterized in that, It also includes the following steps: S5. Obtain meteorological risk data for the work area, determine the warning level based on the meteorological risk data, and generate warning push information for the employee based on the warning level.

8. An artificial intelligence-based personnel safety management system for the express delivery industry, characterized in that, It includes a personnel intelligent training subsystem, a personnel violation supervision subsystem, and an interface adaptation layer that connects the personnel intelligent training subsystem and the personnel violation supervision subsystem; The intelligent personnel training subsystem is configured to: generate training content for the employee based on the employee's job-related data; generate a training effectiveness evaluation result based on the employee's completion of the training content; and grant the employee job access in response to the training effectiveness evaluation result reaching a preset evaluation threshold. The personnel violation supervision subsystem is configured to: obtain the work video of the employee who has the work permission and is in the on-duty work state; perform violation identification on the work video to obtain the violation judgment result; update the employee's violation points according to the violation judgment result; and generate a reinforcement training instruction in response to the violation points meeting the reinforcement training trigger condition. The interface adaptation layer is configured to: transmit the enhanced training instruction from the personnel violation supervision subsystem to the personnel intelligent training subsystem, drive the regeneration and re-execution of the training content for the employee, and lock the employee's work access during the regeneration and re-execution.

9. A computer device, characterized in that, It includes: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to: The method for managing the safety of personnel in the express delivery industry based on artificial intelligence, as described in any one of claims 1 to 7, shall be implemented.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the artificial intelligence-based personnel safety management method for the express delivery industry as described in any one of claims 1 to 7.

Images